Why Your AI Keeps Forgetting Your Tech Stack (And How to Fix It)

Every conversation with Claude starts the same way: "I'm using Next.js 15, TypeScript, shadcn/ui, Prisma..." Sound familiar? I got tired of repeating myself.

The Problem: 20+ Messages Just to Set Context

I'm an indie dev working on a Next.js project. Standard stack: TypeScript strict mode, App Router, shadcn/ui, Prisma. Nothing fancy.

But every time I open a new chat with Claude or any AI assistant, it's like talking to someone with goldfish memory. "What framework?" "Pages Router or App Router?" "What's your styling approach?"

Before I get to my actual question, I've burned 20+ messages on context. That's not efficient—that's repetitive busywork. I hate repetitive busywork.

The Solutions I Tried (And Why They Failed)

Copy-paste prompts

Made a "starter prompt" file. Paste it every conversation. Worked for a week:

• File got out of sync with my project
• Still manual effort every time
• Different AI tools had different quirks

.cursorrules file

Step up from copy-paste. Define stack once, it persists. But:

• Only works in Cursor
• Limited structure
• Switch to Claude Code? Back to square one.

System prompts

Custom instructions in ChatGPT/Claude settings. Better, but:

• Character limits kill detailed configs
• Still platform-specific
• No version control

I needed something that works across Cursor, Claude Code, and any AI tool. Something I could actually maintain.

Agent Skills: One Skill, Works Everywhere

Found the Agent Skills open standard. Dead simple idea: create a SKILL.md file with your context, any compatible AI tool loads it automatically.

Here's mine (simplified):

---
description: "Use when building my Next.js SaaS application"
---

## Tech Stack
- **Framework**: Next.js 15 (App Router only)
- **Language**: TypeScript (strict mode)
- **UI**: shadcn/ui components
- **Database**: Prisma with PostgreSQL
- **Auth**: NextAuth.js v5

## Code Conventions
- Server Components by default
- 'use client' only when needed
- No `any` types
- Zod for all validation

Drop this in your project root. Claude knows you. First response is actually what you needed.

Real Results

After a month with Agent Skills:

• Zero "what framework" questions—context loads from SKILL.md
• ~40% fewer tokens per conversation—less explaining basics
• Works in Cursor AND Claude Code—same file, different tools

Best part? When my stack changes (say, Prisma to Drizzle), update one file. Done. No copy-pasting across platforms.

Getting Started

I use Antigravity Skills to browse and manage skills. They have a curated library—useful if you don't want to write everything from scratch.

Quick setup:

1. Browse the skill library or create your own SKILL.md
2. Copy the file to your project root
3. Open Cursor or Claude Code—context loads automatically

That's it. No config, no API keys, no monthly fee.

What's Your Stack?

What repetitive context do you find yourself explaining to AI tools? Drop a comment. Always looking for patterns I might be missing.

---

This post was written with AI assistance for editing and structure.